A Deterministic Framework for Neural Network Quantum States in Quantum Chemistry.

Area of Science:

• Quantum chemistry
• Computational physics
• Machine learning in science

Background:

• Stochastic optimization methods for neural network quantum states (NQS) suffer from sampling variance and slow convergence.
• Accurate simulation of strongly correlated systems remains a significant challenge in computational quantum chemistry.

Purpose of the Study:

• To develop a deterministic optimization framework for NQS that avoids sampling issues.
• To enable efficient and accurate calculations for large, strongly correlated quantum systems.

Main Methods:

• A deterministic optimization framework is proposed, projecting a neural backflow ansatz onto dynamically evolving configuration subspaces.
• The method optimizes variational components of the wave function and estimates residual correlation using posthoc second-order perturbative correction.
• A hybrid CPU-GPU architecture is employed for efficient computation.

Main Results:

• The framework demonstrates sublinear wall-time scaling with subspace size on hybrid architectures.
• Accurate calculations were performed for systems with up to 10^23 configurations, including the chromium dimer.
• Benchmarks on molecular bond dissociations showed stable convergence and accuracy comparable to reference methods.

Conclusions:

• The deterministic NQS optimization framework offers a robust alternative to stochastic methods.
• This approach facilitates the accurate simulation of challenging strongly correlated quantum systems.
• The method shows promise for advancing computational quantum chemistry and materials science.